Re: [PATCH 2/3] cpu: add a CPU_DOWN_PREPARE notifier

[Nick Piggin <nickpiggin@yahoo.com.au>]

Nick Piggin wrote:
> Rusty Russell wrote:
> 
>> On Wed, 2004-09-08 at 22:52, Nick Piggin wrote:
>>
>>> 2/3
>>>
>>> Rusty, can I do this?
>>>
>>> ______________________________________________________________________
>>> Add a CPU_DOWN_PREPARE hotplug CPU notifier. This is needed so we can
>>> dettach all sched-domains before a CPU goes down, thus we can build
>>> domains from online cpumasks, and not have to check for the possibility
>>> of a CPU coming up or going down.
>>
>>
>>
>> And if taking the CPU down fails?  If you need this, you need the
>> CPU_DOWN_FAILED as well, unfortunately.  Hence I prefer the "do the
>> domain thing while machine is frozen" and sidestep it entirely.
>>
> 
> Really? It doesn't need to be run from the stop_machine_run
> context at all - it can happily be done while the system is
> running.
> 
> That said, if you really object to CPU_DOWN_PREPARE and CPU_DOWN_FAILED,
> it probably shouldn't be too much work. Should it make the call from
> take_cpu_down?
> 

The other thing is, it is actually a lot nicer to have this done while
the machine is running, because then cpu_attach_domain guarantees a
quiescent state, so the attach can be done completely atomically from
the point of view of the rest of the scheduler.

So you can always rely on domains and cpu_online_map staying in synch.
The alternative is more fastpath checking of the cpu_online_map, and
possibly more hotplug locks.

In short, I'd really like to have it done this way.